# Soprano-80M: an 80M Apache 2.0 TTS that claims 2000x real time on GPU

> Soprano-80M is a tiny English-only TTS with under 15 ms streaming latency on GPU, no cloning, Apache 2.0; superseded by Soprano 1.1.

Published 2025-12-22

Canonical: https://shipped.anth.us/open-weights/ekwek-soprano-80m/

Soprano-80M is an English-only text-to-speech model of about 80M parameters, published under the name Eugene Kwek on Hugging Face. The card gives 22 December 2025 as its release date. It claims up to 2000 times real time on GPU and 20 times real time on CPU, streaming latency under 15 ms on GPU and under 250 ms on CPU, under 1 GB of memory, and 32 kHz output. It runs on CUDA, CPU and MPS on Windows, Linux and macOS, with a WebUI, CLI, Python API and OpenAI-compatible endpoint, though the endpoint is non-streaming only.

The card is candid about limits: there is no voice cloning, it was trained on only about 1,000 hours of audio, and numbers and unusual words are sometimes mispronounced. Best results come from sentences of 2 to 15 seconds.

The licence is Apache-2.0. The card also says this model is outdated: Soprano-1.1-80M, released 14 January 2026, claims "95% fewer hallucinations and a 63% preference rate" over it. This item covers the family; the successor is the one to try.

Fast and small, with a card that tells you to use the newer one.

## Sources
- [Eugene Kwek model card on Hugging Face](https://huggingface.co/Eugene Kwek/Soprano-80M)
- [Hugging Face repo metadata (creation date)](https://huggingface.co/api/models/Eugene Kwek/Soprano-80M)

## Editor's first pass (AI-assisted)
- Real 4.5: Apache 2.0 weights with CLI, WebUI, Python API and an OpenAI-compatible endpoint.
- Useful 3: Very fast English narration, but no voice cloning and the card marks it outdated.
- Notable 3: A single-developer model that reached over 300 likes; superseded by Soprano 1.1.
- Novel 3: Speed is the pitch; trained on only about 1,000 hours.